comfyui-mimo-asr

ComfyUI MiMo ASR

A focused ComfyUI custom node adapter for XiaomiMiMo/MiMo-V2.5-ASR. It lets ComfyUI workflows transcribe AUDIO inputs with the official MiMo ASR Python API.

What This Repository Maintains

• A ComfyUI node named MiMo V2.5 ASR.
• Path handling for local MiMo source, model, and audio tokenizer assets.
• Language options for Auto, Chinese, and English recognition.
• Lazy runtime loading so ComfyUI can start without immediately importing the large ASR stack.
• Tests for adapter behavior that do not download or load model weights.

This repository does not redistribute MiMo model weights, tokenizer weights, upstream source code, private audio, transcripts, or generated media.

Upstream Project

MiMo-V2.5-ASR is maintained by XiaomiMiMo:

• Model repository: https://github.com/XiaomiMiMo/MiMo-V2.5-ASR
• Hugging Face model: https://huggingface.co/XiaomiMiMo/MiMo-V2.5-ASR
• Hugging Face Space: https://huggingface.co/spaces/XiaomiMiMo/MiMo-V2.5-ASR

Check the upstream repositories for current model, code, and weight licenses before redistributing any upstream assets.

Install

Clone this adapter into your ComfyUI custom nodes directory:

cd ComfyUI/custom_nodes
git clone https://github.com/gaoqi125/comfyui-mimo-asr.git

Install the official MiMo ASR dependencies according to the upstream instructions. The upstream project currently expects Python 3.12 and CUDA-capable Linux for the reference runtime.

Download the official model and tokenizer outside this repository:

pip install huggingface-hub
hf download XiaomiMiMo/MiMo-Audio-Tokenizer --local-dir ./models/MiMo-Audio-Tokenizer
hf download XiaomiMiMo/MiMo-V2.5-ASR --local-dir ./models/MiMo-V2.5-ASR

Clone or otherwise make the official Python source available, then set the node fields:

• source_root: path to the official MiMo-V2.5-ASR source checkout.
• model_path: path to downloaded MiMo-V2.5-ASR weights.
• tokenizer_path: path to downloaded MiMo-Audio-Tokenizer weights.

You can also set environment variables:

export MIMO_ASR_SOURCE_ROOT=/path/to/MiMo-V2.5-ASR
export MIMO_ASR_MODEL_PATH=/path/to/models/MiMo-V2.5-ASR
export MIMO_AUDIO_TOKENIZER_PATH=/path/to/models/MiMo-Audio-Tokenizer

Restart ComfyUI. The node appears under audio/asr.

Node Inputs

• audio: ComfyUI AUDIO input.
• language: Auto, Chinese, or English.
• source_root: optional official source checkout path.
• model_path: local model path.
• tokenizer_path: local audio tokenizer path.
• cache_model: keep the loaded model in memory between runs.

Use cache_model only when the host has enough GPU or system memory. Leave it disabled for safer one-off runs.

Test

python3 -m unittest discover -s tests -v

The tests validate adapter logic only. They do not require ComfyUI, CUDA, PyTorch, the MiMo model, or the tokenizer.

Maintenance Scope

This adapter is meant to track ComfyUI integration behavior: node metadata, path configuration, runtime loading, sample-rate handling, and safe setup errors. Upstream model accuracy, model architecture, official dependencies, and weight distribution belong to XiaomiMiMo.

Support and Security

Use GitHub Issues for reproducible adapter bugs. Do not attach private audio, model weights, .env files, API keys, or full ComfyUI logs that may contain local paths or personal data. See SECURITY.md and SUPPORT.md.